Bundled models

Models ship inside the package under fiberhmm/models/. --enzyme (and for Hia5 --seq) selects them; -m/--model is only for custom models. For a bundled model, the enzyme/platform registry decides the observation mode even if the file's own metadata disagrees; a custom model uses its embedded mode.

Presets

File	--enzyme	--seq	Mode	Role
hia5_pacbio.json	hia5	pacbio	pacbio-fiber	HMM and TF recall
hia5_nanopore.json	hia5	nanopore	nanopore-fiber	HMM and TF recall (rebuilt from naked-DNA and untreated controls in 3.0)
dddb_nanopore.json	dddb	any	daf	HMM and TF recall (in-vivo accessible-state rates on the reindexed naked table, 3.0)
ddda_nuc.json	ddda	any	daf	first-pass nucleosome HMM
ddda_TF.json	ddda	any	daf	TF recall (model card)
ddda_nuc_refine.json	ddda	any	daf	internal radial-nucleosome likelihoods (context-independent: one hit probability per state)
ddda_nuc_profile.json	ddda	any	—	radial nucleosome profile ddda_phase_posterior_v1 (dyad and helical-phase template)

The public preset list is fiberhmm/models/SUPPORTED_MODES.json (schema fiberhmm.supported_modes.v1); every preset uses tf_min_llr 5.0.

For DddA, fiberhmm-call --enzyme ddda runs all three tables in one pass; keeping them separate means that updating the TF table does not retune the HMM or the radial nucleosome caller. The nucleosome profile's identity and digest are recorded in the output (nuc_model, nuc_sha256).

Other bundled files

File	Used by	Content
ddda_duplex_v1.json	fiberhmm-pair (--call-layer input-ma)	sequence-free duplex ranker for archived-MA calls
ddda_duplex_rotational_v1.json	fiberhmm-pair (--call-layer rotational-recall, the default for 3.0 calls)	the same ranker calibrated on 3.0 rotational DddA calls
ecogii_pacbio.json	custom -m only	EcoGII m6A (development)
cpg_nanopore.json	custom -m only	CpG methyltransferase, cpg mode (development)

The duplex models carry status externally_replicated_experimental and their training and validation intervals.

Legacy tables

Kept under fiberhmm/models/legacy/ only to reproduce older results; never selected by a preset.

File	What it is
hia5_nanopore_gt_swapped_legacy.json	the Nanopore Hia5 table as shipped through 2.x
hia5_nanopore_v2.9_reindexed_legacy.json	the 2.x Nanopore Hia5 values in encoder order, bundled in 3.0 development builds before the control-built table (3.0 development builds only)
dddb_nanopore_gt_swapped_legacy.json	the DddB table as shipped through 2.x
dddb_nanopore_naked_2f10003c.json	the reindexed naked DddB table the in-vivo table is built on (3.0 development builds only)
ddda_nuc_refine_context_v2.6.json	the earlier per-context radial-nucleosome likelihoods (3.0 development builds only)
ddda_pacbio.json	the older one-pass DddA model, superseded by ddda_nuc.json + ddda_TF.json
hia5_pacbio_fp0.1x.json	an experimental low-false-positive Hia5 PacBio calibration that did not generalize

To reproduce a 2.x Nanopore Hia5 call, pass the legacy table explicitly:

LEGACY=$(python -c "import os, fiberhmm.models as m; print(os.path.join(m.__path__[0], 'legacy', 'hia5_nanopore_gt_swapped_legacy.json'))")
fiberhmm-call -i ont.bam -o ont.2x-table.bam -m "$LEGACY" --enzyme hia5 --seq nanopore

Other 3.0 defaults also changed (see Upgrading from 2.x); add the matching options, such as --prob-threshold 128, to reproduce 2.x numbers closely.

A source checkout also has a top-level models/ directory: a compatibility mirror for scripts that used paths such as models/hia5_pacbio.json. The packaged fiberhmm/models/ is authoritative.

Model file format

Models are JSON:

{
  "model_type": "FiberHMM",
  "version": "2.0",
  "n_states": 2,
  "startprob": [...],
  "transmat": [[...], [...]],
  "emissionprob": [[...], [...]],
  "context_size": 3,
  "mode": "pacbio-fiber"
}

Each emission row has, for k = 3, 8,194 entries: 4,096 context codes for "marked", two sentinel positions, and 4,096 for "unmarked". Context codes use the encoder's base order (A = 0, C = 1, T = 2, G = 3; see Training). The loader orders the states protected/accessible from the emission rows. fiberhmm-utils inspect model.json prints a summary.

.npz and legacy .pickle models still load. Loading a pickle executes code: load or fiberhmm-utils convert only pickles you trust, and prefer JSON.